In recent years, importance of security is increased, and an automatic recognizing apparatus such as face recognition or finger print recognition is now becoming widespread. From the viewpoints of ease and safety of people, automatic recognition for various objects (pedestrians or vehicles in the periphery) using a sensor mounted to a vehicle for preventing traffic accidents now attracts the public attention.
In such recognition, an automatic recognition system is achieved from different types of input sensor information by eventually recognizing patterns thereof. In general, the term “patterns” takes the form of “feature vectors” obtained by extracting the features from the input sensor information.
Various multi-variable analyses may be employed in classification of the vectors, and are generally classified into linear classification and non-linear classification. The term “linear classification” indicates classification achieved by applying linear transformation to an entered vector, and the term “non-linear classification” indicates classification achieved by applying non-linear transformation to the vector.
Regarding the linear classification, learning methods employing various statistical methods have been proposed, such as linear discriminant analysis described in L. Chen, H. Liao, M. Ko, J. Lin, and G. Yu, “Anew LDA-based face recognition system which can solve the small sample size problem,” Pattern Recognition, Vol. 33, No. 10, pp. 1713-1726, 2000 and a Support Vector Machine (SVM) described in Christopher J. C Burges, “A Tutorial on Support Vector Machines for Pattern Recognition”, Data Mining and Knowledge Discovery, Vol. 2, No. 2, pp. 121-167, 1988, incorporated by reference.
On the other hand, in the case of the non-linear classification, there are a few effective learning methods since the non-linear transformation cannot be obtained easily. However, Kernel SVM using Kernel method disclosed in “A Tutorial on Support Vector Machines for Pattern Recognition” and Boosting (AdaBoost, Real AdaBoost, Joint Boosting) have produced good results. Boosting includes a plurality of weak classifiers disclosed in Y. Freund and R. E. Schapire, “A decision-theoretic generalization of on-line learning and an application to boosting,” Journal of Computer and System Sciences 55(1), 1997, R. Schapire and Y. Singer, “Improved Boosting Algorithms using confidence-rated predictions,” Machine Learning, Vol. 37, No. 3, 1999, and A Torralba, K Murphy and W Freeman, “Sharing Features: efficient boosting procedures for multiclass object detection,” In Proc. IEEE Conference on Computer Vision and Pattern Recognition, 2004. In Kernel SVM, the non-linear transformation is performed by replacing the inner product of the vectors by Kernel function, and the non-linear property is expressed by effectively combining the weak classifiers.
The classifiers in the related art as described above have advantages and disadvantages as described below.
Since most of the linear classifiers employ an inner product of the vectors, a calculation cost required for classification is low. However, they have a limit that these classifiers are effective only when the distribution of the target vectors is linearly separable. Therefore, they cannot be effective classifiers for the non-linear distributions.
On the other hand, the non-linear classifiers represented by Kernel SVM or Boosting are effective for most of non-linear distributions. However, since calculation of Kernel function and calculation of the weak classifiers are required by number of times in Kernel SVM and in Boosting respectively, the calculation cost is significantly increased.
For example, in video surveillance applications, hardware resources which can process a large amount of calculation, such as a personal computer (PC) or equivalent apparatuses, may be utilized for executing the applications. Therefore, the classifiers such as Kernel SVM or Boosting which has a high classification performance but also requires high calculation cost may be utilized in the above applications.
However, in view of general application, the recognizing apparatus to be mounted to vehicles for example, the hardware which can be installed is limited to the small one having a performance lower than personal computers.
In the environment such as the above-described video surveillance, if the implementation with low-performance hardware is possible, the cost may be reduced correspondingly.
Therefore, it is desirable to use a classifier with low calculation cost. However, in the case of the linear classifier, although the low calculation cost may be achieved, the constraint of “linear” may result in significantly insufficient classification performance.